AI cracks superbug problem in two days that took scientists years

What the AI Actually Did

• The system (Google’s “AI co-scientist”) generated a ranked list of hypotheses for a specific microbiology problem.
• Its top hypothesis matched a conclusion the team had spent years developing; several other hypotheses “made sense,” and one is now being investigated as new.
• The AI did not prove anything experimentally; it suggested lines of inquiry that still require lab verification.

Novelty of the Hypothesis

• Many commenters argue the hypothesis was likely present in some form in prior literature or “future work” sections, or implicit in the team’s earlier publications.
• Follow‑up reporting referenced in the thread says the system was explicitly fed a 2023 paper by the same group; it may simply have ignored a limitation that the humans had assumed.
• Some see this as “when given a decade of work as premises, the computer produced the conclusion in hours,” not discovery from scratch.

Prompting, Suggestion, and “Clever Hans”

• Several note that the original prompt itself contained strong hints (e.g., mentioning the “tail”), making it easier for the model to converge on the desired answer.
• This is compared to Clever Hans or mentalism: the model appears smarter because the human encodes much of the solution in the question.
• Others frame it as “rubber‑ducking”: carefully explaining the problem to the AI helps the human see connections.

Media Hype and Misleading Framing

• The headline “cracks superbug problem in two days” is widely criticized as sensational and inaccurate.
• Commenters emphasize: AI proposed a plausible hypothesis based on existing work; it didn’t independently solve how to kill superbugs.
• Some call the article “PR for Google,” noting the timing with the product launch and the BBC’s lack of technical skepticism.

LLMs as Advanced Search / Synthesis Tools

• Many see this as LLMs shining at literature synthesis and hypothesis generation—“the next generation of search,” not magic reasoning.
• Value is in connecting scattered, published “bits” across papers and disciplines faster than humans can manually.

Data, Privacy, and Training Concerns

• Debate centers on whether the model might have used private data (Gmail, Drive, unpublished drafts).
• Google’s Workspace policy text is dissected; interpretations differ on what “without permission” and “outside of Workspace” really mean.
• There’s broader worry about opaque training sources and over‑attributing capabilities that may just be regurgitated or recombined prior work.

Impact on Science and Society

• Some wonder how such tools will affect the role of grad students and exploratory “messing around” in research.
• Others warn of societal risks in believing AI has capacities (true understanding, originality) it does not demonstrably possess.